Combined annotation exploration for SCPCP000006

Introduction

The aim is to combine label transfer and CNV inference to annotate Wilms tumor samples in SCPCP000006. The proposed annotation will be based on the combination of:

• the label transfer from the fetal kidney reference (Stewart et al.), in particular the fetal_kidney_predicted.compartment and fetal_kidney:predicted.cell_type, as well as the prediction.score for each compartment,

• the predicted CNV calculated using intra-sample endothelial and immune cells (--reference both) as normal reference; no reference was used for samples with fewer than 3 predicted normal cells.

We will then explore and validate the chosen annotation.

We will use some of the markers genes to validate visually the annotations.

The analysis can be summarized as the following:

first level annotation	second level annotation	selection of the cells	marker genes for validation	CNV validation
normal	endothelial	compartment == "endothelium" & predicted.score > pred.thr	VWF	no CNV
normal	immune	compartment == "immune" & predicted.score > pred.thr	PTPRC, CD163, CD68	no CNV
normal	kidney	compartment == "fetal_nephron" & predicted.score > pred.thr & cnv_score < cnv.thr.lower	CDH1, PODXL, LTL	no CNV
normal	stroma	compartment == "stroma" & predicted.score > pred.thr & cnv_score < cnv.thr.upper	VIM	no CNV
cancer	stroma	compartment == "stroma" & cnv_score > cnv.thr.upper	VIM	proportion_cnv_chr: 1, 4, 11, 16, 17, 18
cancer	blastema	compartment == "fetal_nephron" & cell_type == "mesenchymal cell" & cnv_score > cnv.thr.upper	CITED1	proportion_cnv_chr: 1, 4, 11, 16, 17, 18
cancer	epithelial	compartment == "fetal_nephron" & cell_type != "mesenchymal cell" & cnv_score > cnv.thr.upper	CDH1	proportion_cnv_chr: 1, 4, 11, 16, 17, 18
unknown	-	the rest of the cells	-	-

with the defined parameters:

• pred.thr = 0.75
• cnv.thr.lower = 0
• cnv.thr.upper = 2

Packages

library(Seurat)
library(tidyverse)
library(patchwork)
library(DT)

Base directories

# The base path for the OpenScPCA repository, found by its (hidden) .git directory
repository_base <- rprojroot::find_root(rprojroot::is_git_root)

# The current data directory, found within the repository base directory
data_dir <- file.path(repository_base, "data", "current", "SCPCP000006")

# The path to this module
module_base <- file.path(repository_base, "analyses", "cell-type-wilms-tumor-06")

result_dir <- file.path(module_base, "results")

Output files

The report will be saved in the notebook directory. The final annotation file SCPCP000006-annotations.tsv will be saved in the results directory.

# cell type annotation table
annotations_tsv <- file.path(result_dir, "SCPCP000006-annotations.tsv")

Input files

In this notebook, we are working with all of the samples in SCPCP000006. The sample metadata can be found in sample_metadata_file in the data folder.

# sample metadata
sample_metadata_file <- file.path(repository_base, "data", "current", "SCPCP000006", "single_cell_metadata.tsv")
metadata <- read.table(sample_metadata_file, sep = "\t", header = TRUE)

From the pre-processed and labeled Seurat object in results/{sample_id}/06_infercnv_HMM-i3_{sample_id}_reference-{reference}.rds, we will extract the following per-cell information:

• the predicted compartment in fetal_kidney_predicted.compartment and fetal_kidney_predicted.compartment.score
• the predicted cell type in fetal_kidney_predicted.cell_type and fetal_kidney_predicted.cell_type.score
• the sample identifier in sample_id
• the CNV {dupli|loss} estimate per chromosome {i}, where i is in {1-22}, and arm {p|q}: has_{dupli|loss}_chr{i}{p|q}
• the counts of markers genes

# Extract from each Seurat object information that will be used
sample_ids <- metadata |>
  dplyr::filter(seq_unit != "spot") |>
  dplyr::pull(scpca_sample_id) |>
  unique()

# These samples were run with "none" as the reference
none_reference_samples <- c("SCPCS000177", "SCPCS000180", "SCPCS000181", "SCPCS000190", "SCPCS000197")


# Create a data frames of all annotations
cell_type_df <- sample_ids |>
  purrr::map(
    # For each sample_id, do the following:
    \(sample_id) {
      if (sample_id %in% none_reference_samples) {
        reference <- "none"
      } else {
        reference <- "both"
      }

      input_file <- file.path(
        result_dir,
        sample_id,
        glue::glue("06_infercnv_HMM-i3_{sample_id}_reference-{reference}.rds")
      )

      # The file may not be present if this is being run in CI, which is ok.
      # If we are not running in CI and the file doesn't exist, we should error out
      # We should error out if the file does not exist and we are NOT testing
      if (!file.exists(input_file)) {
        if (params$testing) {
          return(NULL)
        } else {
          stop("Input RDS file does not exist.")
        }
      }

      # Read in the Seurat object
      srat <- readRDS(input_file)

      # Create and return a data frame from the Seurat object with relevant annotations
      # this data frame will have four columns: barcode, sample_id, compartment, organ
      data.frame(
        srat@meta.data,

        # cell embedding
        umap = srat@reductions$umap@cell.embeddings,

        # marker genes
        PTPRC = FetchData(object = srat, vars = "ENSG00000081237", layer = "counts"),
        VWF = FetchData(object = srat, vars = "ENSG00000110799", layer = "counts"),
        VIM = FetchData(object = srat, vars = "ENSG00000026025", layer = "counts"),
        CITED1 = FetchData(object = srat, vars = "ENSG00000125931", layer = "counts"),
        CDH1 = FetchData(object = srat, vars = "ENSG00000039068", layer = "counts"),
        PODXL = FetchData(object = srat, vars = "ENSG00000128567", layer = "counts"),
        COL6A3 = FetchData(object = srat, vars = "ENSG00000163359", layer = "counts"),
        SIX2 = FetchData(object = srat, vars = "ENSG00000170577", layer = "counts"),
        NCAM1 = FetchData(object = srat, vars = "ENSG00000149294", layer = "counts"),
        THY1 = FetchData(object = srat, vars = "ENSG00000154096", layer = "counts")
      ) |>
        tibble::rownames_to_column("barcode") |>
        dplyr::mutate(sample_id = sample_id)
    }
  ) |>
  # now combine all dataframes to make one big one
  dplyr::bind_rows()

We will also use:

• the samples metadata or clinical data reported in /Users/sjspielman/ALSF/open-scpca/OpenScPCA-analysis/data/current/SCPCP000006/single_cell_metadata.tsv
• the cytoband file to get the length of each chromosome arm (for plotting only)

# Add clinical metadata
cell_type_df <- cell_type_df |>
  left_join(metadata, by = c("sample_id" = "scpca_sample_id"))

# the cytoband file that will be used to extract the size of each chromosome arm for plotting only
arm_order_file <- file.path(module_base, "results", "references", "hg38_cytoBand.txt")
# Load cytoBand file
cytoBand <- read.table(gzfile(arm_order_file), header = FALSE, sep = "\t", stringsAsFactors = FALSE)
colnames(cytoBand) <- c("chrom", "start", "end", "band", "stain")
# Here we define the size of each chromosome arm that will be used for plotting of the mean CNV profile
# Extract chromosome arm information
cytoBand <- cytoBand |>
  mutate(arm = substr(band, 1, 1)) # Extract 'p' or 'q' from the band column

# Define arm boundaries
chromosome_arms <- cytoBand |>
  group_by(chrom, arm) |>
  summarize(
    Start = min(start),
    End = max(end),
    .groups = "drop"
  ) |>
  mutate(
    Size = End - Start,
    dupli = paste0("has_dupli_", chrom, arm),
    loss = paste0("has_loss_", chrom, arm),
    chrom_arm = glue::glue("{chrom}{arm}")
  ) |>
  mutate(Size = Size / sum(Size)) |>
  # Define chromosome arm order
  mutate(chrom_arm = factor(chrom_arm, levels = c(
    paste0("chr", rep(1:22, each = 2), c("p", "q")),
    "chrXp", "chrXq", "chrYp", "chrYq"
  ))) |>
  # Sort genes by Chromosome arm and Start position
  arrange(chrom_arm, Start) |>
  filter(arm != "")

Functions

do_Feature_mean

do_Feature_mean shows heatmap of mean expression of a feature grouped by a metadata.

• df is the name of the table containing metadata and feature expression (counts) per cells
• group.by is the name of the metadata to group the cells
• feature is the name of the gene to average the expression

do_Feature_mean <- function(df, group.by, feature) {
  df <- df |>
    group_by(sample_id, !!sym(group.by)) |>
    summarise(m = mean(!!sym(feature)))

  p <- ggplot(df, aes(x = sample_id, y = !!sym(group.by), fill = m)) +
    geom_tile() +
    scale_fill_viridis_c() +
    theme_bw() +
    theme(text = element_text(size = 20)) +
    theme(axis.text.x = element_text(angle = 90, hjust = 0.5, vjust = 0.5)) +
    guides(fill = guide_colourbar(title = paste0(feature)))

  return(p)
}

do_BoxPlot_mean

do_BoxPlot_mean shows boxplot of the percentage of cells with CNV in a specific chromosome arm per clinical relevant group (metadata)

• df is the name of the table containing metadata and feature expression (counts) per cells
• group.by is the name of the metadata to group the cells
• comparison is a list of factors in group.by for which we will compare the means using an unpaired Wilcoxon test

do_BoxPlot_mean <- function(df, group.by, comparison) {
  tmp <- df |>
    select(!!sym(group.by), sample_id, contains("has_loss_chr"), contains("has_dupli_chr")) |>
    group_by(!!sym(group.by), sample_id) |>
    summarise_all(.funs = mean) |>
    pivot_longer(cols = -c(sample_id, !!sym(group.by))) |>
    rename("chromosome_arm" = "name", "percentage" = "value") |>
    mutate(chromosome_arm = factor(chromosome_arm, levels = c(has_dupli_chr, has_loss_chr)))

  p <- ggplot(tmp, aes(x = !!sym(group.by), y = percentage, fill = !!sym(group.by))) +
    geom_boxplot() +
    scale_fill_brewer(palette = "PuRd") +
    geom_dotplot(binaxis = "y", stackdir = "center") +
    facet_wrap(~chromosome_arm, nrow = 5) +
    theme_classic()

  return(p)
}

Analysis

Global CNV score

As done in 06_cnv_infercnv_exploration.Rmd, we calculate single CNV score and assess its potential in identifying cells with CNV versus normal cells without CNV.

For each cell, we checked the level of CNV for each chromosome.

• If the cell has less than 0 chromosome with CNV, the global cnv_score will be FALSE.
• If the cell has more than 2 chromosome with CNV, the global cnv_score will be TRUE.
• The default is set to unknown.

The infercnv method can lead to false positive CNV. We introduced a flexibility of +2 over the cnv_threshold to reduce the risk of misclassification of normal cells as cancer cells.

Please note that some cancer cell might not have any CNV. There is thus a risk of misclassifying cancer cells as normal cells. However, we cannot avoid this risk with the data and tools currently available. This is a limitation of our annotations.

cell_type_df <- cell_type_df |>
  mutate(cnv_score = rowSums(cell_type_df[, grepl("has_loss_chr|has_dupli_chr", colnames(cell_type_df))], na.rm = TRUE)) |>
  mutate(cnv_score = case_when(
    cnv_score > params$cnv_threshold_high ~ "CNV",
    cnv_score <= params$cnv_threshold_low ~ "no CNV",
    .default = "unknown"
  ))

First level annotation

At first, we like to indicate in the first.level_annotation if a cell is normal, cancer or unknown.

• normal cells can be observed in all five compartments (normal,endothelium, immune, stroma or fetal nephron) and do not have CNV.
• cancer cells are either from the stroma or fetal nephron compartments and must have at least few CNV

We only allow a bit of flexibility in terms of CNV profile for immune and endothelium cells that have a high predicted score. Indeed, we know that false positive CNV can be observed in a cell type specific manner. Please note that in 06_infercnv.R, we renamed all immune and endothelium cells with a fetal_kidney_predicted.compartment.score > 0.75 as normal.

The threshold used for the predicted.score is defined as a parameter of this notebook as 0.75. The thresholds used for the identification of CNV are also defined as notebook parameters with a minimum of 0 and a maximum of 2.

# Define normal cells
# We first pick up the immune and endothelial cells annotated via the label transfer compartments under the condition that the predicted score is above the threshold
cell_type_df <- cell_type_df |>
  mutate(first.level_annotation = case_when(
    # assign normal/cancer based on condition
    fetal_kidney_predicted.compartment %in% c("fetal_nephron", "stroma") &
      cnv_score == "no CNV" ~ "normal",
    fetal_kidney_predicted.compartment %in% c("normal", "immune", "endothelium") &
      fetal_kidney_predicted.compartment.score > params$predicted.celltype.threshold ~ "normal",
    fetal_kidney_predicted.compartment %in% c("fetal_nephron", "stroma") &
      cnv_score == "CNV" ~ "cancer",
    .default = "unknown"
  ))

Using this basic strategy, we identified 157520 cancer cells, 13557 normal cells.

29460 cells remain unknown.

ggplot(cell_type_df, aes(x = umap.umap_1, y = umap.umap_2, color = first.level_annotation)) +
  geom_point(size = 0.5, shape = 19, alpha = 0.5) +
  facet_wrap(facets = ~sample_id, ncol = 5) +
  theme_bw() +
  theme(text = element_text(size = 22))

cell_type_df |>
  dplyr::count(first.level_annotation, sample_id) |>
  tidyr::pivot_wider(
    names_from = first.level_annotation,
    values_from = n
  ) |>
  DT::datatable()

Strikingly, 5 samples show more normal cells than cancer cells:

• SCPCS000177
• SCPCS000180
• SCPCS000181
• SCPCS000190
• SCPCS000197

These five samples do not have enough immune and endothelium cell to be used as a normal reference in scripts/06_infercnv.R and we previously decided to run scripts/06_infercnv.R without any reference for these samples. Without a specified reference, the mean expression profile across the sample is taken as the normal reference, biasing the inference of CNV. For those samples, the annotations based on infercnv cannot be trusted. To avoid any confusion in the following steps of the analysis, we decided to force the annotation of the entire samples to unknown.

cell_type_df$first.level_annotation[cell_type_df$sample_id %in% none_reference_samples] <- "unknown"

Second level annotation

Normal cells

• Normal cells from the fetal nephron compartment must be normal kidney cells.
• Normal cells from the stroma compartment must be normal stroma cells.
• Immune and endothelial cells have been already identified by label transfer.

Cancer cells

Wilms tumor cancer cells can be:

• cancer stroma: We define as cancer stroma all cancer cells from the stroma compartment.
• blastema: We defined as blastema every cancer cell that has a fetal_kidney_predicted.cell_type == mesenchymal cell. We know that these mesenchymal cells are cells from the cap mesenchyme that are not expected to be in a mature kidney. These blastema cells should express higher CITED1 and/or NCAM1.
• cancer epithelium: We defined as cancer epithelium all cancer cells that are neither stroma nor blastemal cells. We expect these cells to express epithelial markers. Their predicted cell type should correspond to more mature kidney epithelial subunits.

# Define a vector of `fetal_kidney_predicted.cell_type` levels that corresponds to immune cells
immune_celltypes <- c(
  "B cell",
  "CD4-positive, alpha-beta T cell",
  "conventional dendritic cell",
  "lymphocyte",
  "macrophage",
  "mast cell",
  "monocyte",
  "natural killer cell",
  "neutrophil",
  "plasmacytoid dendritic cell"
)



cell_type_df <- cell_type_df |>
  mutate(second.level_annotation = case_when(
    # assign normal cells based on condition
    (sample_id %in% none_reference_samples) ~ "unknown",
    (cell_type_df$fetal_kidney_predicted.compartment == c("fetal_nephron") &
      cell_type_df$cnv_score == "no CNV") ~ "kidney",
    (cell_type_df$fetal_kidney_predicted.compartment == c("stroma") &
      cell_type_df$cnv_score == "no CNV") ~ "normal stroma",
    (cell_type_df$fetal_kidney_predicted.compartment == c("normal") &
      cell_type_df$fetal_kidney_predicted.cell_type == c("endothelial cell")) ~ "endothelium",
    (cell_type_df$fetal_kidney_predicted.compartment == c("normal") &
      cell_type_df$fetal_kidney_predicted.cell_type %in% immune_celltypes) ~ "immune",

    # assign cancer cells based on condition
    (cell_type_df$fetal_kidney_predicted.compartment == c("stroma") &
      cell_type_df$cnv_score == "CNV") ~ "cancer stroma",
    (cell_type_df$fetal_kidney_predicted.compartment == c("fetal_nephron") &
      cell_type_df$cnv_score == "CNV" &
      cell_type_df$fetal_kidney_predicted.cell_type == "mesenchymal cell") ~ "blastema",
    (cell_type_df$fetal_kidney_predicted.compartment == c("fetal_nephron") &
      cell_type_df$cnv_score == "CNV" &
      cell_type_df$fetal_kidney_predicted.cell_type != "mesenchymal cell") ~ "cancer epithelium",
    .default = "unknown"
  ))

Second.level_annotation of cancer and normal cells - umap reduction

ggplot(cell_type_df, aes(x = umap.umap_1, y = umap.umap_2, color = second.level_annotation)) +
  geom_point(size = 0.5, shape = 19, alpha = 0.5) +
  facet_wrap(facets = ~sample_id, ncol = 5) +
  theme_bw() +
  theme(text = element_text(size = 22))

Second.level_annotation of cancer and normal cells without unknown - umap reduction

cell_type_df_sub <- cell_type_df[cell_type_df$second.level_annotation != "unknown", ]
ggplot(cell_type_df_sub, aes(x = umap.umap_1, y = umap.umap_2, color = second.level_annotation)) +
  geom_point(size = 0.5, shape = 19, alpha = 0.5) +
  facet_wrap(facets = ~sample_id, ncol = 5) +
  theme_bw() +
  theme(text = element_text(size = 22))

Per sample, the annotations in the umap reduction seem to make sense as we can the three main Wilms tumor components are easily distinguished:

• cancer blastema
• epithelium
• stroma

Validation cancer versus normal based on the CNV profile

We look for each second.level_annotation and sample_id at the proportion of cells that has CNV in each of the chromosome.

These heatmaps allow us to:

• validate that mostly tumor cells do present high percentages of cells with CNV (by definition)
• identify chromosome and cell types that might be more sensitive to FALSE positive CNV, such as chr6p which is a hub for immune genes
• get the repartition per sample of the CNV

Loss on long arms

# create a list of all plots by mapping over the vector of indices, instead of a for loop
plot_arm <- ifelse(params$testing, "", "q") # only use "q" if NOT testing.

1:22 |>
  purrr::map(
    \(i)
    do_Feature_mean(cell_type_df_sub, group.by = "second.level_annotation", feature = glue::glue("has_loss_chr", i, plot_arm))
  ) |>
  # send into patchwork to make a single plot from the list of plots in 2 columns
  patchwork::wrap_plots(ncol = 2)

gains on long arms

# create a list of all plots by mapping over the vector of indices, instead of a for loop
plot_arm <- ifelse(params$testing, "", "q") # only use "q" if NOT testing.
1:22 |>
  purrr::map(
    \(i)
    do_Feature_mean(cell_type_df_sub, group.by = "second.level_annotation", feature = glue::glue("has_dupli_chr", i, plot_arm))
  ) |>
  # send into patchwork to make a single plot from the list of plots in 2 columns
  patchwork::wrap_plots(ncol = 2)

loss on short arms

# create a list of all plots by mapping over the vector of indices, instead of a for loop
plot_arm <- ifelse(params$testing, "", "p") # only use "p" if NOT testing.
c(1:12, 16:20) |>
  purrr::map(
    \(i)
    do_Feature_mean(cell_type_df_sub, group.by = "second.level_annotation", feature = glue::glue("has_loss_chr", i, plot_arm))
  ) |>
  # send into patchwork to make a single plot from the list of plots in 2 columns
  patchwork::wrap_plots(ncol = 2)

gains on short arms

# create a list of all plots by mapping over the vector of indices, instead of a for loop
plot_arm <- ifelse(params$testing, "", "p") # only use "p" if NOT testing.
c(1:12, 16:20) |>
  purrr::map(
    \(i)
    do_Feature_mean(cell_type_df_sub, group.by = "second.level_annotation", feature = glue::glue("has_dupli_chr", i, plot_arm))
  ) |>
  # send into patchwork to make a single plot from the list of plots in 2 columns
  patchwork::wrap_plots(ncol = 2)

Mean CNV profile

While the above heatmaps can be informative, because there are so many of them it is difficult to get a global picture of the CNV profile of the entire dataset.

Here we aim to plot a mean CNV profile across the 35 annotated samples as presented in Figure 1A of Cresswell et al. (2024). This represents the mean SNP Array profile of 64 pediatric kidney cancer (including 60 Wilms tumors).

Here we plot for each chromosome arm (x-axis) the number of samples having a dupli (y-axis > 0, in red) or a loss (y-axis < 0, in blue).

For a sample, a CNV is defined as clonal (opaque) if detected in more than 70% of the cells; and subclonal (transparent) when detected in at least 10% of the cells.

# Here we define the vector of ordered chromosomes arms that will be used for plotting
has_dupli_chr <- names(cell_type_df)[stringr::str_starts(names(cell_type_df), "has_dupli_chr")]

# define the vector of ordered chromosomes arms
has_loss_chr <- names(cell_type_df)[stringr::str_starts(names(cell_type_df), "has_loss_chr")]

# prepare the table for the number of sample with a (sub)clonal loss

CNV_profile_loss <- cell_type_df_sub |>
  select(sample_id, contains("has_loss_chr")) |>
  group_by(sample_id) |>
  summarise_all(.funs = mean) |>
  pivot_longer(
    cols = -sample_id,
    names_to = "chromosome_arm",
    values_to = "percentage"
  ) |>
  left_join(chromosome_arms, by = c("chromosome_arm" = "loss")) |>
  mutate(chromosome_arm = factor(chromosome_arm, levels = has_loss_chr)) |>
  mutate(
    absolute_70 = ifelse(percentage > 0.7, -1, 0),
    absolute_10 = ifelse(percentage > 0.1, -1, 0)
  )


# prepare the table for the number of sample with a (sub)clonal gain

CNV_profile_gain <- cell_type_df_sub |>
  select(sample_id, contains("has_dupli_chr")) |>
  group_by(sample_id) |>
  summarise_all(.funs = mean) |>
  pivot_longer(
    cols = -sample_id,
    names_to = "chromosome_arm",
    values_to = "percentage"
  ) |>
  left_join(chromosome_arms, by = c("chromosome_arm" = "dupli")) |>
  mutate(chromosome_arm = factor(chromosome_arm, levels = has_dupli_chr)) |>
  mutate(
    absolute_70 = ifelse(percentage > 0.7, 1, 0),
    absolute_10 = ifelse(percentage > 0.1, 1, 0)
  )


# plot
# Here we chose to loop on chromosome arms to then be able to scale each bar to the length of the corresponding genomic region using wrap_plot.
# Of note, the resulting plot is way better than the one obtained with facet_wrap or facet_grid so far.

loss_profile <- list()
for (chr in unique(CNV_profile_loss$chromosome_arm)) {
  loss_profile[[chr]] <- ggplot(
    CNV_profile_loss[CNV_profile_loss$chromosome_arm == chr, ],
    aes(x = chromosome_arm, y = absolute_10)
  ) +
    geom_bar(stat = "identity", fill = "#41b6c4", alpha = 0.4) +
    geom_bar(aes(x = chromosome_arm, y = absolute_70), stat = "identity", fill = "#41b6c4") +
    scale_x_discrete(labels = gsub("has_loss_", "", chr)) +
    theme_void() +
    ylim(c(-40, 0)) +
    theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust = 1, size = 12), text = element_text(size = 15))
}
loss_profile <- wrap_plots(loss_profile, nrow = 1, widths = chromosome_arms$Size[chromosome_arms$loss %in% unique(CNV_profile_loss$chromosome_arm)], guides = "collect")

gain_profile <- list()
for (chr in unique(CNV_profile_gain$chromosome_arm)) {
  gain_profile[[chr]] <- ggplot(CNV_profile_gain[CNV_profile_gain$chromosome_arm == chr, ], aes(x = chromosome_arm, y = absolute_10)) +
    geom_bar(stat = "identity", fill = "#ce1256", alpha = 0.4) +
    geom_bar(aes(x = chromosome_arm, y = absolute_70), stat = "identity", fill = "#ce1256") +
    theme_void() +
    ylim(c(0, 40)) +
    theme(axis.text.x = element_blank())
}
gain_profile <- wrap_plots(gain_profile, nrow = 1, widths = chromosome_arms$Size[chromosome_arms$dupli %in% unique(CNV_profile_gain$chromosome_arm)], guides = "collect")


# combined plot
wrap_plots(gain_profile, loss_profile, ncol = 1) + plot_layout(widths = CNV_profile_gain$Size)

Comparing with the profile in Cresswell et al. (2024), we identified important CNV known to play a role in Wilms tumor, including:

• chr8q, chr12, chr13, chr18q gains
• chr11q, chr16, chr17q loss

However, we spotted some differences:

• We couldn’t detect any chr4q loss as described in Cresswell et al. (2024).

• We identified 2 losses in chr5 and chr6 that do not show up in Cresswell et al. (2024). These may be infercnv induced false-positive CNV.

Our major concern is the pattern of the chr1. We do see a tendency of chr1p loss and chr1q gain but it is not as pronounced as in Cresswell et al. (2024). chr1q gain is however one of the most promising predictive marker of Wilms tumor outcome (Nelson et al. (2020)). It is however important to notice that the profile shown in Cresswell et al. (2024) derived from SIOP samples that have been pre-treated with chemotherapy, while samples in the dataset are enriched in upfront resection samples that haven’t been pre-treated, according to the COG protocol. So far, we do not know how the pre-operative chemotherapy impact the clonal selection of specific CNV.

Mean of the percentage of cells with CNV in a specific chromosome arm per clinical relevant group

vital_status

As described in Nelson et al. (2020), Phelps et al. (2018) and Zheng et al. (2023), LOH at chr1p, chr11p and chr16q as well as chr1q gain are prognostic factors.

Here we aim to see if we identify CNV associated with poorer prognosis.

do_BoxPlot_mean(df = cell_type_df_sub, group.by = "vital_status", comparison = c("Alive", "Expired"))

In our study, we validate the association of chr1q gain, chr11q loss and chr14q loss with poorer prognosis.

treatment

As we don’t know the effect of the treatment on the selection of cells with CNV, it is difficult to compare our CNV profile with the one in Cresswell et al. (2024). We those check the repartition of CNV in samples that have been pre-treated (i.e Resection post chemotherapy) with samples that have been removed surgically before any treatment (i.e Upfront resection).

do_BoxPlot_mean(df = cell_type_df_sub, group.by = "treatment", comparison = c("Upfront resection", "Resection post chemotherapy"))

Our analysis suggest that there is indeed an enrichment in cells with CNV after chemotherapy, including a tendency of more chr1q gain samples after chemotherapy. This could explain some of the differences observed between our mean CNV profile and the one published in Cresswell et al. (2024).

Validation of second level annotation using marker genes

Finally, we aim to validate the second level annotation using marker genes.

Immune, PTPRC expression

do_Feature_mean(cell_type_df_sub, group.by = "second.level_annotation", feature = "ENSG00000081237")

Immune cells are well annotated based on PTPRC expression.

Endothelium, VWF expression

do_Feature_mean(cell_type_df_sub, group.by = "second.level_annotation", feature = "ENSG00000110799")

Endothelial cells are well annotated based on VWF expression.

Stroma

(*) To the best of our knowledge, there is no Wilms tumor specific marker of stroma cells. They should express in some extend similar markers as normal stromal cells including VIM, THY1 and COL6A3.

Vimentin expression

do_Feature_mean(cell_type_df_sub, group.by = "second.level_annotation", feature = "ENSG00000026025")

VIM expression does not seem to be either specific or universal.

COL6A3 expression

do_Feature_mean(cell_type_df_sub, group.by = "second.level_annotation", feature = "ENSG00000163359")

COL6A3 expression is higher in Wilms tumor stroma cells and expressed across a wider range of samples.

**THY1* expression

do_Feature_mean(cell_type_df_sub, group.by = "second.level_annotation", feature = "ENSG00000154096")

THY1 expression does not seem to be either specific or universal.

Blastema

(*) To the best of our knowledge, there is no Wilms tumor specific and universal (i.e. expressed by all cells in every sample) marker of blastema cells. We expect however blastema cells to express higher levels of stemness markers such as CITED1, SIX1/2, NCAM1.

CITED1 expression

do_Feature_mean(cell_type_df_sub, group.by = "second.level_annotation", feature = "ENSG00000125931")

Blastema cells are enriched in CITED1+ cells, but some samples do not express it at all.

NCAM1 expression

do_Feature_mean(cell_type_df_sub, group.by = "second.level_annotation", feature = "ENSG00000149294")

NCAM1 seems to be more cancer specific and broadly expressed among samples.

SIX2 expression

do_Feature_mean(cell_type_df_sub, group.by = "second.level_annotation", feature = "ENSG00000170577")

Epithelium

CDH1 expression

do_Feature_mean(cell_type_df_sub, group.by = "second.level_annotation", feature = "ENSG00000039068")

As expected, epithelial cells express higher expression of CDH1.

PODXL expression

do_Feature_mean(cell_type_df_sub, group.by = "second.level_annotation", feature = "ENSG00000128567")

Create annotation table for export

This section creates the cell type annotation table for export.

annotations_table <- cell_type_df |>
  dplyr::select(
    cell_barcode = barcode,
    scpca_sample_id = sample_id,
    tumor_cell_classification = first.level_annotation,
    cell_type_assignment = second.level_annotation
  ) |>
  mutate(
    # change cancer --> tumor, but keep the other labels
    tumor_cell_classification = ifelse(
      tumor_cell_classification == "cancer", "tumor", tumor_cell_classification
    ),
    cell_type_assignment = str_replace_all(
      cell_type_assignment,
      "cancer ",
      "tumor "
    )
  )

write_tsv(annotations_table, annotations_tsv)

Confirm how many samples we have annotations for:

length(unique(annotations_table$scpca_sample_id))

## [1] 40

Conclusion

• Combining label transfer and CNV inference we have produced draft annotations for 35/40 Wilms tumor samples in SCPCP000006. We decided not to annotate samples for which the infercnv couldn’t be run with a normal reference, as we don’t trust the results in that case and want to avoid misclassification (SCPCS000177, SCPCS000180, SCPCS000181, SCPCS000190, SCPCS000197).

• The mean CNV profile of the Wilms tumor cohort seems reasonable in regards to the one published in Cresswell et al. (2024). We identify some commonalities but also some divergence. They can be due to technical limitation of our CNV inference but can also reflect some specificities of this cohort (e.g. pre/post treatment samples). It would be really useful to compare for some sample the infercnv result to a ground truth such as SNP Array (to be requested).

• The heatmaps of CNV proportion and marker genes support our annotations, but signals with some marker genes are very low. Also, there is no universal marker for each entity of Wilms tumor that cover all tumor cells from all patient. This makes the validation of the annotations quite difficult.

• However, we could try to take the problem from the other side, and used the current annotation to perform differential expression analysis and try to find marker genes that are consistent across patient and Wilms tumor histologies.

• In each histology (i.e. epithelial and stroma), the distinction between cancer and non cancer cell is difficult (as expected). In this analysis, we suggested to rely on the CNV score to assess the normality of the cell. Here again, we could try to run differential expression analysis and compare epithelial (resp. stroma) cancer versus non-cancer cells across patient, aiming to find a share transcriptional program allowing the classification cancer versus normal.

• In our annotation, we haven’t taken into account the favorable/anaplastic status of the sample. However, as anaplasia can occur in every (but do not has to) Wilms tumor histology, I am not sure how to integrate the information into the annotation.

• This notebook could be finally rendered using different parameters, i.e. threshold for the CNV score and predicted score to use.

Session Info

# record the versions of the packages used in this analysis and other environment information
sessionInfo()

## R version 4.4.0 (2024-04-24)
## Platform: aarch64-apple-darwin20
## Running under: macOS 15.3.2
## 
## Matrix products: default
## BLAS:   /Library/Frameworks/R.framework/Versions/4.4-arm64/Resources/lib/libRblas.0.dylib 
## LAPACK: /Library/Frameworks/R.framework/Versions/4.4-arm64/Resources/lib/libRlapack.dylib;  LAPACK version 3.12.0
## 
## locale:
## [1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8
## 
## time zone: America/New_York
## tzcode source: internal
## 
## attached base packages:
## [1] stats     graphics  grDevices utils     datasets  methods   base     
## 
## other attached packages:
##  [1] DT_0.33            patchwork_1.2.0    lubridate_1.9.3    forcats_1.0.0     
##  [5] stringr_1.5.1      dplyr_1.1.4        purrr_1.0.2        readr_2.1.5       
##  [9] tidyr_1.3.1        tibble_3.2.1       ggplot2_3.5.1      tidyverse_2.0.0   
## [13] Seurat_5.1.0       SeuratObject_5.0.2 sp_2.1-4          
## 
## loaded via a namespace (and not attached):
##   [1] RColorBrewer_1.1-3     rstudioapi_0.16.0      jsonlite_1.8.8        
##   [4] magrittr_2.0.3         spatstat.utils_3.1-0   farver_2.1.2          
##   [7] rmarkdown_2.28         vctrs_0.6.5            ROCR_1.0-11           
##  [10] spatstat.explore_3.3-2 htmltools_0.5.8.1      sass_0.4.9            
##  [13] sctransform_0.4.1      parallelly_1.38.0      KernSmooth_2.23-24    
##  [16] bslib_0.8.0            htmlwidgets_1.6.4      ica_1.0-3             
##  [19] plyr_1.8.9             plotly_4.10.4          zoo_1.8-12            
##  [22] cachem_1.1.0           igraph_2.0.3           mime_0.12             
##  [25] lifecycle_1.0.4        pkgconfig_2.0.3        Matrix_1.7-0          
##  [28] R6_2.5.1               fastmap_1.2.0          fitdistrplus_1.2-1    
##  [31] future_1.34.0          shiny_1.9.1            digest_0.6.37         
##  [34] colorspace_2.1-1       rprojroot_2.0.4        tensor_1.5            
##  [37] RSpectra_0.16-2        irlba_2.3.5.1          crosstalk_1.2.1       
##  [40] labeling_0.4.3         progressr_0.14.0       fansi_1.0.6           
##  [43] spatstat.sparse_3.1-0  timechange_0.3.0       httr_1.4.7            
##  [46] polyclip_1.10-7        abind_1.4-5            compiler_4.4.0        
##  [49] bit64_4.0.5            withr_3.0.1            fastDummies_1.7.4     
##  [52] highr_0.11             MASS_7.3-61            tools_4.4.0           
##  [55] lmtest_0.9-40          httpuv_1.6.15          future.apply_1.11.2   
##  [58] goftest_1.2-3          glue_1.7.0             nlme_3.1-166          
##  [61] promises_1.3.0         grid_4.4.0             Rtsne_0.17            
##  [64] cluster_2.1.6          reshape2_1.4.4         generics_0.1.3        
##  [67] gtable_0.3.5           spatstat.data_3.1-2    tzdb_0.4.0            
##  [70] data.table_1.16.0      hms_1.1.3              utf8_1.2.4            
##  [73] spatstat.geom_3.3-2    RcppAnnoy_0.0.22       ggrepel_0.9.5         
##  [76] RANN_2.6.2             pillar_1.9.0           vroom_1.6.5           
##  [79] spam_2.10-0            RcppHNSW_0.6.0         later_1.3.2           
##  [82] splines_4.4.0          lattice_0.22-6         bit_4.0.5             
##  [85] survival_3.7-0         deldir_2.0-4           tidyselect_1.2.1      
##  [88] miniUI_0.1.1.1         pbapply_1.7-2          knitr_1.48            
##  [91] gridExtra_2.3          scattermore_1.2        xfun_0.47             
##  [94] matrixStats_1.3.0      stringi_1.8.4          lazyeval_0.2.2        
##  [97] yaml_2.3.10            evaluate_0.24.0        codetools_0.2-20      
## [100] cli_3.6.3              uwot_0.2.2             xtable_1.8-4          
## [103] reticulate_1.38.0      munsell_0.5.1          jquerylib_0.1.4       
## [106] Rcpp_1.0.14            globals_0.16.3         spatstat.random_3.3-1 
## [109] png_0.1-8              spatstat.univar_3.0-0  parallel_4.4.0        
## [112] dotCall64_1.1-1        listenv_0.9.1          viridisLite_0.4.2     
## [115] scales_1.3.0           ggridges_0.5.6         crayon_1.5.3          
## [118] leiden_0.4.3.1         rlang_1.1.4            cowplot_1.1.3